statistics 1b interludes most studies crave ‘significance’ · posted on aoril 21 2013 126...

3
1 Statistics 1B Interludes 13. ‘Not statistically significant’ Most studies crave ‘significance’ XKCD, Jan 2015 Collects real phrases from academic papers Today’s story! Measured insecticide metabolites in urine of 571 pregnant women 6 years later measured metabolites in 287 children Correlated with behavioural problems 5 metabolites at 3 levels, mothers/children, 3 outcome scales = 60 95% confidence intervals for associations (adjusted with logistic regression) Only one excluded 1.

Upload: others

Post on 16-Sep-2020

1 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Statistics 1B Interludes Most studies crave ‘significance’ · Posted on Aoril 21 2013 126 Comments ..and this is wheæ we put the non-significant tæsults. som - - cards user

1

Statistics 1B Interludes

13. ‘Not statistically significant’

Most studies crave ‘significance’

XKCD, Jan 2015

Collects real phrases from academic papers

Today’s story!• Measured insecticide metabolites in urine of 571

pregnant women• 6 years later measured metabolites in 287 children• Correlated with behavioural problems• 5 metabolites at 3 levels, mothers/children, 3 outcome

scales• = 60 95% confidence intervals for associations

(adjusted with logistic regression)

• Only one excluded 1.

Page 2: Statistics 1B Interludes Most studies crave ‘significance’ · Posted on Aoril 21 2013 126 Comments ..and this is wheæ we put the non-significant tæsults. som - - cards user

2

30 more sensitive tests? Children shown below• Paper and press release only reported

the few significant results [2 positive and 1 negative]

• A green jelly-bean example?

• And maybe children with behaviouralproblems get more head lice? [reverse causation]

But sometimes ‘non-significance’ is of interest …

February 2015

But did it really show no benefit?

Page 3: Statistics 1B Interludes Most studies crave ‘significance’ · Posted on Aoril 21 2013 126 Comments ..and this is wheæ we put the non-significant tæsults. som - - cards user

3

Conclusions

• Point estimates for all consumption levels show protection

• Confidence intervals are wide as few deaths in the baseline (never-drinker) category

• Wide CIs include plausible protective effects• But authors essentially interpret ‘not

significantly different’ as ‘no effect’• A serious misuse of statistics